The Search: The Effect of the College

Scorecard on Interest in Colleges

Nick Huntington-Kleina

November 16, 2016

Abstract
The College Scorecard is a website launched by the U.S. Depart-
ment of Education in September 2015 that provides information about
different colleges. This paper studies the effects of the website on in-
terest in colleges, calculating both intent-to-treat and a local average
treatment effect estimates of the effect of the College Scorecard. In-
terest is measured using Google search activity. The Scorecard led
to more searches for keywords associated with high-earnings, high-
graduation rate, and low-tuition colleges. However, the size of the
effect is very small.

Keywords: college choice, policy evaluation, information

JEL: I22, I23, I28, D83

a
Corresponding Author. Department of Economics, California State University
Fullerton, 800 N. State College Blvd., Fullerton, CA 92831-3599. nhuntington-
klein@fullerton.edu.

1
1 Introduction
When making decisions about higher education, students and their families
must learn what colleges are out there and what they are like, before applying
to and attending a given college. Ideally, they will accurately understand the
consequences of their decision. However, a lot of information about a given
college is unavailable except through a thorough investigation, likely above
the heads of many high schoolers.
Attempting to address this problem is the College Scorecard.1 Announced
in 2013 and fully launched in September 2015, the College Scorecard is a
website compiling detailed information on a long and nearly comprehensive
list of American colleges and universities. The site lists commonly available
information like each college’s location and enrollment size. Using the federal
government’s financial aid database, the Scorecard adds information that
was previously difficult or impossible to find, such as average net tuition for
financial aid recipients, student loan repayment rates, and the earnings of
graduates.
Part of the purpose of the College Scorecard is to assist in funding de-
cisions for higher education, and indeed the project was originally launched
as a replacement for the Obama administration’s controversial plan to rank
colleges. The provision of information to students and families is also a ma-
jor stated goal of the website. As emphasized by the Secretary of Education
Arne Duncan at the time the project was announced (Feb. 13, 2013), “We
know students and families are often overwhelmed in the college search pro-
cess but feel they lack the tools to sort through the information and decide
which school is right for them [...] The College Scorecard provides a snapshot
about an institutions cost and value to help families make smart decisions
about where to enroll.” The site offers a huge amount of data to potential
students, was widely publicized at the time of launch, and is free to use.
The Scorecard is an intervention that attempts to change student be-
havior by providing information to students. Interventions of this sort have
been tested a number of times in research contexts, and are often of interest
because they promise behavioral response that, even if it is small, comes at
a very low cost per treatment. The Scorecard may be the largest example of
a real-world implementation of such a policy in education.
The goal of this study is to examine the degree to which the release of
1
https://collegescorecard.ed.gov/

2
the College Scorecard data affected interest in different colleges. I examine
whether colleges with high marks in the attributes prominently advertised by
the Scorecard (post-graduation earnings, net tuition, and graduation rate)
saw increases in their Google search activity as a result of the Scorecard.
I produce an estimate of the effect of the Scorecard on aggregate Google
search behavior, as measured by Google Trends. Google Trends reports the
intensity of search behavior on Google for given keywords over time, relative
to the pool of all Google searches. There is a fair amount of work using
Google Trends as a tool that allows for instantaneous predictions of cur-
rent variables like the unemployment rate Choi and Varian (2012). However,
Google Trends data also has the capacity to be useful in observational esti-
mation of casual effects, when search activity is a reasonable measure of the
variable of interest. Stephens-Davidowitz (2014) uses Google Trends data to
estimate the casual effect of racial animus on Barack Obama’s vote share.
Racial animus is measured by Google Trends results for racial epithets, to
get around the hesitancy to report racist attitudes in survey data.
Google search behavior acts as a proxy for “first-step” interest in colleges,
and is capable of picking up additional curiosity about a given college even if
students eventually decide not to apply for any reason. Using IPEDS data, I
find that in past years, a Google Trends index one unit higher is associated
with 57 more applications for the college, or 3.3 more applications when
controlling for college and year fixed effects.
The Scorecard had statistically significant effects on Google search activ-
ity, and those effects were of the expected signs. Colleges with high post-
graduation earnings, high graduation rates, and low tuition saw increases
in their search behavior relative to other colleges as a result of the Score-
card. However, these effect sizes were extremely small. Colleges saw bumps
in their search activity that was .5%, .8%, 5.6% of a standard deviation of
search activity relative to other colleges with tuition $1,000 higher, median
earnings $1,000 lower, or graduation rates 5% lower, respectively. Effects
were stronger for colleges that primarily offer bachelor’s degrees than for
colleges that primarily offer associate’s degrees.
These estimates of the College Scorecard are intent-to-treat estimates,
and do not take into account that most people performing searches for col-
leges likely did not visit the College Scorecard (or may not be students). I
apply a method for estimating the local average treatment effect (LATE) in
aggregate data. This method uses the implementation of the College Score-
card policy as an instrument for the Google Trend Index for the Scorecard

3
itself. The LATE estimate is also small. Even comparing colleges that are
very different in terms of their Scorecard-reported attributes, the effect of
adding of a single treated user on aggregate search results is on the order of
one ten-millionth of a standard deviation of search activity.
This paper finds that the College Scorecard did have an effect on search
behavior. However, this effect is small enough to be considered negligible.
The policy may be justifiable given that its marginal costs of operation are
likely low, and the information has meaningful uses for researchers and gov-
ernment policy makers. However, the site is unlikely to have large effects on
student behavior as intended.

2 Literature
This study looks into how student interest in different colleges relates to the
attributes of those colleges, and how that interest changes in response to in-
formation revelation. Both of these topics have long and thorough literatures
generally. There has been a recent boom in studies examining the intersec-
tion between educational choice and the revelation of information, and I will
focus on that intersection here.
The decisions of whether or not to go to college, which college to go to,
and what to major in are important determinants of individual and aggre-
gate human capital. Multiple considerations drive these decisions. Benefits
include the immediate consumption value of education (Wiswall and Zafar,
2015a; Jacob et al., 2016) as well as long-term financial and non-financial ben-
efits such as improved labor market performance, improved health outcomes,
and improved marriage market performance (Oreopoulos and Salvanes, 2011;
Oreopoulos and Petronijevic, 2013). The costs of higher education include
tuition and the possibility of being burdened with student loans (Dynarski,
2003).
The literature on college choice is not short of evidence that each of these
incentives play some part in college choice, although results vary on the
weight each is given in the decision-making process.
A growing literature shows that students are not fully aware of exactly
how large these costs and benefits are. Mostly following from Dominitz and
Manski (1996), there has been considerable interest in eliciting student beliefs
about the labor market outcomes associated with different education levels,
different institutions, or different majors (among many others, Kodde, 1987;

4
Smith and Powell, 1990; Betts, 1996; Avery and Kane, 2004; Rouse, 2004;
Botelho and Pinto, 2004; Attanasio and Kaufmann, 2009; Zafar, 2011; At-
tanasio and Kaufmann, 2012; Stinebrickner and Stinebrickner, 2014; Hastings
et al., 2016; Huntington-Klein, 2015, 2016b) as well as student beliefs about
tuition (Avery and Kane, 2004; Usher, 2005; Grodsky and Jones, 2007; Booij
et al., 2012).
A general summation of this literature is that students have relatively
noisy beliefs about both labor market outcomes and the costs associated with
college, although beliefs are on average not wildly inaccurate. These beliefs
are associated with student choice. Students choose educational options they
expect to have higher labor market returns, and are more likely to go to
college if they think tuition is low.
The inaccuracy of student beliefs about the costs and benefits of college
in general, or different college majors in particular, is a cause for concern.
Students necessarily must make decisions on the basis of their perceptions of
costs and benefits, rather than the truth (Manski, 1993). Without accurate
beliefs it seems unlikely that students would be able to make well-considered
decisions about their education. Further, students from disadvantaged back-
grounds tend to have less accurate beliefs (Avery and Kane, 2004; Rouse,
2004; Hastings et al., 2016; Huntington-Klein, 2016b). Disparities in educa-
tional attainment may have to do with differences in beliefs and informational
access.
The upside of undesirable educational choices following from student be-
liefs is that it opens a potentially inexpensive policy lever for improving
educational choices. Following from the literature on student beliefs has
been a string of studies that attempt to alter student beliefs by presenting
them with new information. Ideally, these altered beliefs will then change
matriculation patterns.
A good example of this type of intervention is in Wiswall and Zafar
(2015b). The authors surveyed undergraduate students at New York Univer-
sity. These students were asked to report their beliefs about the earnings of
current workers in the labor force conditional on having completed particu-
lar college majors, as well as their expectations for what their own earnings
would be having graduated with that same major. Then, students were pro-
vided with information about the actual earnings of current workers, with a
goal of providing publicly available information that students might be able
to access themselves. Finally, students reported their beliefs again, so any
revisions could be observed.

5
 Wiswall and Zafar (2015b) find that students do meaningfully revise their
beliefs in response to the new information. There are large changes in beliefs
reported after the interventions. Then, when beliefs are elicited again two
years later, beliefs correlate more strongly with post-intervention beliefs than
pre-intervention beliefs. However, despite the changes in beliefs, Wiswall and
Zafar (2015a) find that the changes in actual behavior that arise from the
changes in beliefs are small relative to other influences. This result can be
interpreted as meaning that students do not weight earnings heavily in their
decisions, or that the revision in reported beliefs does not translate into a
revision in the actual beliefs that students base decisions on.
The result that informational interventions have relatively small behav-
ioral effects, even when reported beliefs change, is not limited to work by
Wiswall & Zafar. Many intervention studies that intend to give information
to students as a means to change beliefs and thus behavior find behavioral
effects that are small or localized within particular subgroups, even in cases
where beliefs are changed (Oreopoulos and Dunn, 2013; Kerr et al., 2015;
Hastings et al., 2015; Bergman et al., 2016; Fryer, 2017) Bettinger et al.
(2012) perform a study in which they provide information and aid about
the FAFSA to customers of tax preparation service H&R Block. They find
that, while interventions offering direct assistance in filling out the FAFSA
forms did have meaningful and large effects on submitting the FAFSA and
matriculating in college, an intervention offering information alone did not
have similar effects.
Information-based educational interventions have had some success. Nguyen
(2008) and Jensen (2010) find considerable response to information about
the returns to secondary education in Madagascar and the Dominican Re-
public, respectively, although the context and level of education are very
different than in studies about college. Perhaps the biggest success story re-
lating to college is in Hoxby and Turner (2013), in which the authors target
high-achieving, low-income students with information about their ability to
attend and pay for selective colleges. While they test several interventions,
including some that make the application process easier, they find that even
information alone has meaningful effects on applications and matriculation.2
2
There is also the growing related literature on “nudges” from short messages that
remind the recipient about certain tasks they should be doing, such as in Castleman and
Page (2015), which often find meaningful effects on behavior, especially given the low cost
of the interventions. However, with the exception of messages that intend to inform (as in
Fryer, 2017), rather than remind, these interventions should not be understood as purely

6
 These generally mixed results from informationally based interventions
does not mean that students are not open to the use of information at all. One
source of publicly available and well-known information about college quality
is the U.S. News and World Report, which ranks colleges on several attributes
and provides overviews to the public. There is evidence that information
about college quality as portrayed in the U.S. News and World Report affects
college choice, at least for some kinds of students or colleges (Parker and
Summers, 1993; Monks and Ehrenberg, 1999; Buss et al., 2004). Changes in
rankings from year to year, which may be likely to overstate actual changes
in quality and thus may be taken as largely an informational effect, influence
the applications that students send (Griffith and Rask, 2007).
The College Scorecard falls somewhere between the publicly available
general college overview and ranking available in the U.S. News and World
Report and the careful, directed information revelation of the experimental
literature.
Like the U.S. News and World Report, it is publicly available. However,
the Scorecard is less well-established and less well-known. A Google Trends
comparison of the search terms “US News College Rankings” and “College
Scorecard” shows that the U.S. News search term beats the Scorecard by
about two to one since the launch of the Scorecard. This likely understates
the difference in popularity, since “US News College Rankings” is a very
specific search term. As such, we might expect the Scorecard to have less
effect than the U.S. News and World Report.
Like the experimental studies, the College Scorecard focuses on informa-
tion that is generally not publicly available, or is difficult to find (earnings
in particular). Also, the site is made available with the express purpose of
correcting student beliefs and improving decision-making. The mixed results
of these studies suggest that it is unlikely the College Scorecard will have
a major effect, especially since the Scorecard is exposed only to those who
seek it out, as opposed to an experiment in which all subjects at least see
the information.
These expectations are validated in Hurwitz and Smith (2016), which ex-
amines whether student behavior in sending SAT scores to particular schools
changed after the introduction of the College Scorecard. They find that
tuition and graduation rate information did not have an effect on student
SAT-sending behavior. However, they find that a 10% increase in the earn-
informational.

7
ings associated with a college increases the number of SAT scores received
by 2.4%, an increase driven by students with advantaged backgrounds.
This study adds to this literature on the effects of information revelation
on student choice. Clearly, it is most similar to Hurwitz and Smith (2016),
which examines the effects of the same policy. This paper measures the effect
of the information on Google search activity. Compared to the SAT score
sending in Hurwitz and Smith (2016), Google search activity is a less direct
measure of student matriculation in several ways. Some students may not use
Google to look up colleges, perhaps clicking links directly from the Scorecard
itself every time they want to learn about the college. Additionally, people
other than students can search for college terms (the author, for one), and
so an unknown portion of the search activity represents non-students.
While Google search activity is an indirect measure of student interest, it
also allows the effect of the Scorecard to be captured at colleges that do not
request SAT scores, like most two-year colleges. The addition of two-year
colleges expands the analysis to consider the effect on a wider range of insti-
tutions. Google searches also include responses from students whose interest
is piqued by the new information but do not end up applying, because they
found other aspects unappealing or because they do not believe they have
any chance of being admitted. The effect in this paper should be understood
as the effect of the Scorecard information on interest in colleges, which may
be a more clean estimate of the upper bound of the effects of this sort of
similarly well known publicly-available college information. This paper can
be paired with Hurwitz and Smith (2016) which provides an estimate that is
closer to the effects on matriculation.

3 Data
This study uses two sources of data: the data from the College Scorecard
website on the college attributes reported to visitors, and data from Google
Trends reporting the intensity of search in the United States for search terms
associated with particular colleges.
College Scorecard data includes information on 3,595 two- and four-year
colleges. Most of the Scorecard data are compiled from the Integrated Post-
secondary Education Data System (IPEDS), and the rest comes from federal
databases on financial aid recipients. The website itself includes a means of
searching for colleges by name, region, or type.

8
 Individual college pages include extensive information on each college, but
the most prominent are the enrollment of the college, its location, whether it
is public or private, and three variables that are more difficult to find from
other sources: the graduation rate (a four-year graduation rate for primar-
ily two-year colleges, and a six-year graduation rate for primarily four-year
colleges), the average in-state net price among students who receive federal
financial aid, and the average earnings ten years after graduation among stu-
dents who receive federal financial aid. This information is displayed relative
to national averages. Figure 1 shows how the data are presented to visitors.
This study focuses on these three prominently displayed variables, al-
though much more information is available in dropdown menus.
For each college campus in the College Scorecard sample, I generate a list
of keyword searches for each college campus. Keywords are first generated
using an automatic algorithm that takes as inputs the name of the college and
the college website URL as listed in the College Scorecard data. For example,
one might be “University of California - Santa Cruz” and “www.ucsc.edu/”
The first keyword is the name of the full name of the college, “University
of California Santa Cruz”3 and the second is the base URL of the college
stripped of extraneous detail, “ucsc.edu”. The third drops the domain type,
which generally leaves a common nickname for the college, here “ucsc”. Next,
the algorithm generates abbreviations of the college name, alternately includ-
ing and stripping terms like “of,” “the,” and “at” and alternately abbreviat-
ing terms like “university” as “u,” “community college” as “cc,” or “technical
institute” as “tech” (so we get “U of California Santa Cruz” and “U Cali-
fornia Santa Cruz”), and alternately including or excluding the word “uni-
versity” or “college” when it follows “state” (so “Bismarck State College”
generates “Bismarck State” as a search term). The algorithm also gener-
ates keywords taking into account several widely used abbreviations, such
as “UC” for “University of California” or “PSU” for “Pennsylvania State
University”.
After the algorithm is run, the keyword list is expanded by hand, plug-
ging in likely search terms that are be suggested by the algorithm, like
“Texas Lutheran” for “Texas Lutheran University.” All search terms are
then checked by hand, plugging them into Google one at a time. Search
terms that do not bring up the college in the first five results are dropped.
For example, “kc” is a search term generated by the algorithm for Ketter-
3
Google Trends is not case-sensitive.

9
ing College from the URL “kc.edu” but it is deleted because searching for
“kc” produces only references to Kansas City in the first five results, and not
Kettering College.
I use the Python package pyGTrends to download Google Trends index
information for each search term. The data include a weekly Google Trends
index number for each search term from the beginning of April 2013 to the
end of March 2016. Several keyword or keyword comparison searches gener-
ate no data because the search terms are too rare. Google does not report
Trends results for these terms or comparisons because of privacy concerns.
This leads to some colleges being dropped because all of their keywords re-
turned no results, and so there are 3,439 college campuses in the sample.
The resulting data set links Google Trends indices for each search term
to the College Scorecard information reported for the related college. Google
Trends data does not report direct search volume, but rather an index of
the popularity of a given search term relative to all Google searches, and
so are not directly comparable across search terms. Rather, they report the
number of searches for a term in a given week relative to other weeks covered
by the index and the total number of Google searches in that time period
(Stephens-Davidowitz and Varian, 2014). As such, in analysis it is necessary
to account for a time trend, since the number of other Google searches the
index is relative to changes over time.
Accounting for the time trend allows indices to be compared. This does
not, however, adjust for differences in scale between the indices. A one-unit
change in the index for an extremely popular keyword might represent 10,000
searches, where that same change for an unpopular keyword might represent
10. Results should be interpreted in terms of relative popularity change,
rather than changes in absolute search volume, without further assumptions.
The resulting sample of Google Trends indices is large, with over 1.5 mil-
lion raw keyword-by-week observations, which are collapsed to over 125,000
college-months for most analyses, averaging over each keyword for a given
college. Aggregating all the keywords for a given college into one observation
ensures that colleges with more keywords are not overrepresented, and makes
results easier to interpret. However, since the index for each term is, as above
described, on a different scale, an average of the keywords is not the same
thing as a true Google Trends index for the college itself. In Section 5.1 I
use alternate means of aggregating keywords together.
Summary statistics for the sample are in Table 1. Except for the Google
Trends index, each average is taken at the college level. On average, each

10
 Table 1: Summary Statistics

Variable Mean (SD) Min Max

Earnings $37,601 ($11,446) $11,600 $166,200
Net Tuition $16,687 ($7,833) $393 $45,774
Graduation Rate .437 (.209) .022 1
Enrollment above 15,000 .067
Enrollment 2,000 - 15,000 .385
Enrollment below 2,000 .548
Predominantly Associate’s .420
Predominantly Bachelor’s .580
Index 47.162 (21.497) 0 100
Keywords per College 2.901 (1.324) 1 9
Observation Level N
Colleges 3,439
Keywords 9,976
College-months 127,206
Weekly Index Observations 1,561,739

college has 2.9 keywords associated with it, although some campuses (in
particular those that allow a lot of permutations of “University/U”, “col-
lege”/omitted, and “of the”/omitted) have up to nine. Earnings are the
median earnings of federal financial aid recipients ten years after gradua-
tion, and the mean over campuses of median earnings is $37,601. Similarly,
average net tuition for federal aid recipients is $16,687. The average grad-
uation rate is .437, which includes both the 58% of campuses that are pri-
marily Bachelor’s-granting institutions as well as the 42% that are primarily
Associate’s-granting. More than half of the colleges in the sample are small
and have enrollments below 2,000.
These data can be used to estimate the effect of the college scorecard
using a difference-in-difference specification that is capable of taking into
account the context of the data.

11
Figure 1: College Scorecard Example

12
4 Estimation
In this section I describe the model used to estimate the effect of the College
Scorecard on student interest in colleges, and the assumptions necessary to
identify the estimate.
The model takes into account several features of the data and the inter-
vention. First, the College Scorecard is implemented at a single point in time.
So, the effect must be identified using cross-sectional variation in college at-
tributes likely to interact with the effect of the Scorecard. Second, the effect
of the Scorecard must be separated from long-term trends in the popularity
of colleges on the basis of those attributes. Third, since popularity of a given
search term in Google Trends data is given relative to popularity of the same
term at other times, and not relative to other terms, care must be taken to
scale inter-college comparisons appropriately.
To address all three of these issues, I use a difference-in-difference estima-
tor with continuous college-attribute treatments, time trends, attribute-trend
controls, and college fixed effects. The use of a difference-in-difference esti-
mator with continuous college-attribute treatments identifies the model on
the basis of college attribute differences in the effect of the Scorecard on
interest in a particular college.
A college with low tuition, for example, might be expected to see a greater
increase in Google activity relative to other colleges after the Scorecard is
implemented. Attribute-trend controls address the possibility that these at-
tributes have become more popular over time without the help of the Score-
card. If students have generally been more interested in low-tuition colleges
over time, then the attribute-trend control allows the before/after effect of
the scorecard to be separated from the general trend. Trend controls and
attribute-trend controls also account for changes in the Google Trends data
over time, since the pool of searches that the Google Trends indices are taken
relative to also changes in size over time. College fixed effects allow for the
Google Trends data to be compared across colleges.
In sum, the model of interest is

Indexit = αi + β1 Xi × Scorecardt + β2 Xi × t + β3 Xi × t2
+ β4 t + β5 t2 + β6 Scorecardt + εit (1)
Where Indexit is the aggregated Google Trends index for college i in pe-
riod t, αi is the college fixed effect, Xi is the attribute or attributes of interest,

13
which could include tuition, earnings, or graduation rate, and Scorecardt is
an indicator that the observation is from after the College Scorecard was
implemented. β2 and β3 allow Xi to lead to additional interest over time
regardless of the Scorecard. β̂1 is the estimate of interest, and shows how the
introduction of the Scorecard led to additional interest in colleges with high
levels of attribute Xi .
The trend controls make the observations comparable across time. Google
Trends data provides the popularity of a given search term relative to the
volume of other Google searches, and since the volume of Google searches
changes over time, observations are not directly comparable. The time trend
is intended to adjust for this. Results are robust to the interaction of the time
trend with Scorecardt , essentially forming a regression discontinuity design.
The model is estimated using OLS with robust standard errors.
Indexit combines the Google Trends data for multiple search keywords for
the same college, so that colleges with more keywords are not overweighted.
There are many ways to combine the searches for multiple terms into a single
index, several of which will be covered later for robustness. The main method,
however, uses equal weights, simply averaging together the indices for all
keywords with enough search activity to not be omitted from the data to
produce Indexit .
Time is taken at the month level, adjusted so that each month starts on
the 12th, so that the policy launch date of September 12th, 2015, is not in the
middle of a month. The weekly index score of each search term is averaged
to create Indexit . Results are robust to the use of weekly data instead.
So, in the main model, the observation level is college-by-month, which
is aggregated from keyword-by-week data.
The coefficient of interest in this model is β̂1 , on the interaction term be-
tween an attribute and the post-Scorecard indicator. Given the nonstandard
data source, interpretation is not obvious. β̂1 can be interpreted as the dif-
ference in how much that college’s Google Trends index changed as a result
of the Scorecard being introduced between two colleges with different levels
of the given attribute. So if β̂1 = .5 and the attribute in question is Earnings
in thousands of dollars, then we would say that the Scorecard increased the
Google Trends index for a college with a $20,000 median earnings by .5 more
than it did for a college with a $19,000 median earnings.
Additionally, the effect estimated here should be interpreted as the effect
of the availability of information, not necessarily the effect of beliefs being
changed. Without further data on beliefs there is no way to distinguish when

14
the revelation of Scorecard information actually changes beliefs, as opposed
to revealing something that everyone already knows.
Indexit variable records aggregate search activity. This search activity
includes both non-students searching for colleges, and students who have
never seen the Scorecard. The estimate β̂1 should be properly understood as
the intent-to-treat effect of the Scorecard on aggregate search behavior. In
understanding the effect of the policy on individual behavior, a local average
treatment effect (LATE) will be preferable to intent-to-treat.
To estimate the LATE, I use an instrumental variables method. This
method is similar to Imbens and Angrist (1994), who show that in a ran-
domized experiment with imperfect compliance, standard analysis identifies
an intent-to-treat estimate, and the LATE can be identified using the ran-
domization as an instrument for the receipt of the treatment. In Huntington-
Klein (2016a) I modify the Imbens & Angrist method to apply to aggregate
data in a difference-in-difference context, and use a version of that estimator
here.
Treatment rate is not observed since there is no individual-level data.
Instead of individual-level compliance data, I use an aggregate measure of
interest in the policy. Specifically, data on Google searches for the College
Scorecard. The main model then becomes

Indexit = αi + (β1 N )Xi × ScorecardIndext + β2 t + β3 t2 + εit (2)

Where ScorecardIndext is the index of Google search activity for the

College Scorecard in period t, which stands in for a variable indicating actual
treatment, and N is the approximate number of College Scorecard users each
unit of ScorecardIndext represents, calculated using an observation of the
number of website hits in April-May 2016 at the College Scorecard website.4
The coefficient on Xi × ScorecardIndext is β1 N , which can be adjusted by
N to get β1 , interpretable as “one additional user of the Scorecard website
leads to a β1 Xi increase in the search index for that college.”
The use of the Scorecard and similar information sites, measured in
ScorecardIndext , is of course endogenous, since people select themselves
into treatment. As such, the Scorecardt variable is used as an instrument,
under the assumption that the release of the Scorecard policy affected college
4
https://analytics.usa.gov/ reports website hits in the past 30 days at time of
viewing, but does not report old data on website hits.

15
search behavior only through the actual policy (or interest in similar infor-
mation sources), and that Google searches for the policy is proportional to
actual use. The first stage of the instrumental variables model is then

Xi × ScorecardIndext = α + γ1 Scorecardt + γ2 t + γ3 t2 + εit (3)

The coefficient γ1 captures only the increase in interest that occurs as

a result of the Scorecard policy itself. That is, it isolates the compliers.
The estimate of β1 in Equation 2 is taken by dividing the coefficient on
Xi ×ScorecardInterestt by N , and gives the average effect of the application
of the policy to one additional treated person. Here, β̂1 is the LATE estimate,
with the outcome variable measured on an aggregate level.
This method departs in several ways from Imbens and Angrist (1994).
The identifying variation is not the result of a randomized controlled exper-
iment, data is not at the individual level, and ScorecardIndext is a measure
of interest rather than a direct measure of the receipt of treatment.

5 Results
Table 2 displays the results of four different models of the form presented in
Equation 1.
The coefficients on the Scorecard interaction terms can be taken as the
aggregate treatment effect of the introduction of the College Scorecard on
search activity for colleges with particular characteristics. So, for example,
the .105 coefficient in model (2) indicates that if College A’s graduates earn
$1,000 more than College B’s, then the introduction of the Scorecard raised
the Google Trends index by .105 more for College A than for College B.
The first result that jumps out is the unintuitive sign on Tuition × Score-
card in Model (1). Higher tuition colleges saw more of a bump from the
Scorecard than low-tuition colleges. However, this appears to be due to the
correlation between tuition and the positive attributes of earnings and grad-
uation rate. When all three are included in model (4), the sign becomes
negative.
Earnings and graduation rate have intuitive and expected signs in all
models they appear in. Colleges with higher earnings and higher graduation
rates saw a larger increase in popularity from the Scorecard. The Scorecard
had statistically significant impacts of largely the expected signs.

16
 Table 2: Effects of the College Scorecard on Search Activity for Colleges

Google Search Index

Variable (1) (2) (3) (4)
Tuition ($1k) × .103*** -.040**
Scorecard (.015) (.019)
Earnings ($1k) × .105*** .066***
Scorecard (.010) (.014)
Grad. Rate × 9.879*** 9.453***
Scorecard (.581) (.782)
College Fixed Effects Yes Yes Yes Yes
Quadratic Trend Yes Yes Yes Yes
Quadratic Trend × Yes Yes Yes Yes
Attributes
N 121,582 117,142 113,664 107,115
*/**/*** indicates statistical significance at the 10%/5%/1%
level.

However, while statistically significant, the coefficient sizes are extremely

small. As shown in Table 1, the standard deviation across all index ob-
servations is 21.497. Even within colleges, the minimum, mean, and max-
imum standard deviation of the index is 3.033, 8.461, and 27.041, respec-
tively. As such, using model 4, these results suggest that comparing two
otherwise similar colleges, one of which has tuition $1,000 lower, the low-
tuition college would get an increase in their search index from the Scorecard
that is .040/8.461 = .005 of a standard deviation of search activity. Sim-
ilarly, a college with median earnings $1,000 higher would get a boost of
.066/8.461 = .008, and a college with a graduation rate 5 percentage points
higher would get a boost of .05 × 9.453/8.461 = .056 of a standard deviation.
While the Scorecard does have an effect on aggregate search behavior, that
effect is meaningfully small. Meaningful differences in colleges translate to
very small amounts of differential impact of the Scorecard.
The primary analysis here uses an aggregation of multiple keywords per
college that equally weights each keyword, and treats each college similarly
no matter its size. Alternate approaches are shown in 5.1. The analysis
so far assumes that the average effect of the Scorecard is constant across all
college types, which may be untrue, especially if those students who are likely
to be interested in certain kinds of colleges are also more likely to use the
Scorecard. I evaluate this possibility in Section 5.2. Additionally, while the
effect of the Scorecard is meaningfully small, that the Scorecard has an effect

17
at all may be considered an impressive feat given that only a small fraction
of people searching for colleges are likely to have used it. This is addressed
in Section 5.3.

5.1 Variable and Model Robustness

The Google Search Index for each college used for the models in Table 2
equally weights the Google Trends result from each keyword associated with
the college, and treats a one-unit change in the index as equally meaningful
across all colleges.5
In this section I show results for the same four models in 2 using three al-
ternate calculations of the Google Search index. Using a comparative Google
Trends index from December 2015 to March 2016, which allows up to five
keywords to be compared in popularity against each other, I generate a pop-
ularity weight for each keyword based on how often it is used compared to
other keywords for the same college.6
I use these popularity weights to create two alternate Google Search In-
dex calculations. The first, “Weighted Google Search Index,” constructs a
weighted average of the keywords, using the popularity weights rather than
weighting each keyword equally, and “Most Popular Keyword Search Index”
uses only the Google Trends result for the single most popular keyword as-
sociated with each college.
These two alternate calculations still treat a one-unit change in the index
as equally meaningful across all colleges. However, since each Google Trends
result is indexed from 0 to 100, no matter how popular an individual keyword
is, a single unit change on a very popular search term like “Harvard” might
translate to a change of many more actual searches than a single unit change
on a less popular term. To adjust for this, I take the most popular keyword
for each college (as used for the Most Popular Keyword Search Index) and
use Google Trends to construct a between-college ranking of the popularity
of each college’s most popular keyword.7 I then weight the Most Popular
5
While these results are not shown, results are also robust to the use of logarithms of
earnings and tuition rather than levels, or using different standard specifications of the
time trend.
6
For colleges with more than five keywords, multiple comparative searches were per-
formed, with one overlapping keyword so that all terms could be compared.
7
Since Google Trends does not allow more than five keywords to be compared, I ran-
domly create five-keyword sets and use Google Trends to rank the popularity of each

18
Keyword Search Index by the popularity ranking so that the resulting “Pop-
ularity Scaled Search Index” is still indexed from 0 to 100.
The Popularity Scaled Search Index, like the other search indices, does
not offer enough information to say how many searches represent one unit
on the scale. But it does allow me to test whether the original results arise
only because of the difference in the meaning of the scale across colleges.
Table 3 repeats Table 2 using these three alternate outcome variables.
While point estimates are sensitive to which variable is used, none of them
contradict any of the qualitative findings of Table 2 in which estimates were
not of an economically meaningful size, even though they were, with the
exception of tuition, significant and of expected sign.

5.2 Results by College Type

Table 4 divides the sample in three different ways: first by enrollment size,
then by primary degree awarded, and finally dividing colleges into terciles
based on the median SAT of their incoming classes, or ACT if they do not
request the SAT.8 A large number of colleges do not report median incoming
SAT or ACT in the College Scorecard data, and are dropped for this third
analysis. All analyses mimic model (4) from Table 2, using the unweighted
index and including all three characteristics at the same time.
The division by enrollment size displays some interesting differences. In
particular, effects seem to be generally larger for the largest colleges, espe-
cially for the graduation rate effect. This is particularly interesting since
these analyses use the equal-weight Google Search Index, and so a one-unit
increase for a large college represents many more searches than would a one-
unit increase at a small college. However, even at large colleges the effect
size of each input, with perhaps the exception of graduation rate, is fairly
small.
keyword in the set. I repeat this three times so each keyword is ranked in three different
sets, for a total of (5 − 1) × 3 = 12 “more popular/less popular” comparisons for each key-
word. I then use these 12 comparisons per keyword to rank the popularity of all keywords.
Because there is some noise in the Google Trends results due to sampling, this ranking is
not exact, but is very close to exact.
8
Median SAT is constructed by averaging the nonmissing median math, writing, and
verbal scores (to allow for colleges that report math and verbal but not writing), and
creating a tercile of the score. Then, the process is repeated for colleges that report ACT.
The ACT tercile is used for colleges that do not report SAT.

19
Table 3: Alternate Google Search Index calculations

Weighted Google Search Index

Variable (1) (2) (3) (4)
Tuition ($1k) × .053*** .003
Scorecard (.009) (.012)
Earnings ($1k) × .047*** .029***
Scorecard (.006) (.010)
Grad. Rate × 4.496*** 3.719***
Scorecard (.349) (.547)
N 121,286 116,809 113,442 106,967
Most Popular Keyword Search Index
Variable (1) (2) (3) (4)
Tuition ($1k) × .135*** -.040
Scorecard (.019) (.024)
Earnings ($1k) × .149*** .111***
Scorecard (.013) (.018)
Grad. Rate × 12.224*** 1.462***
Scorecard (.746) (1.008)
N 120,139 115,736 112,443 106,079
Popularity Scaled Search Index
Variable (1) (2) (3) (4)
Tuition ($1k) × .065*** .002
Scorecard (.011) (.014)
Earnings ($1k) × .065*** .051***
Scorecard (.008) (.012)
Grad. Rate × 4.794*** 3.484***
Scorecard (.435) (.627)
N 120,139 115,736 112,443 106,079
All models include college fixed effects, quadratic time trends,
and quadratic trends interacted with all attributes. */**/***
indicates statistical significance at the 10%/5%/1% level.

20
Table 4: Effects of the College Scorecard by College Type

Enrollment < 2,000 2,000-15,000 >15,000

Tuition ($1k) × -.039 .043 .023
Scorecard (.029) (.031) (.085)
Earnings ($1k) × .030 .053** .128*
Scorecard (.020) (.025) (.073)
Grad. Rate × 7.937*** 7.156*** 14.954***
Scorecard (1.036) (1.532) (2.974)
Observations 52,096 46,694 8,325
Median SAT Lowest Middle Highest
Tuition ($1k) × -.050 -.126** -.124***
Scorecard (.061) (.057) (.041)
Earnings ($1k) × .061 .033 .036
Scorecard (.063) (.049) (.029)
Grad. Rate × 5.137 7.833** 2.907
Scorecard (3.289) (3.203) (2.578)
Observations 16,280 17,131 16,132
Pct. Receiving Pell Lowest Middle Highest
Tuition ($1k) × .031 -.004 -.098**
Scorecard (.029) (.035) (.041)
Earnings ($1k) × -.012 .102*** -.045
Scorecard (.022) (.032) (.029)
Grad. Rate × 11.820*** 10.499*** 1.527
Scorecard (1.466) (1.664) (1.574)
Observations 38,332 37,666 31,117
Primary Degree AA BA
Tuition ($1k) × -.019 -.129***
Scorecard (.030) (.025)
Earnings ($1k) × .012 -.032*
Scorecard (.035) (.018)
Grad. Rate × 2.973** 11.618***
Scorecard (1.367) (.999)
N 45,288 61,827
All models include college fixed effects, quadratic time
trends, and quadratic trends interacted with all at-
tributes. */**/*** indicates statistical significance at the
10%/5%/1% level.

21
 The division of colleges by incoming SAT tercile shows that the gradua-
tion rate has the largest effect size for middle-tercile colleges, and the smallest
effect for top-tercile colleges. Of course, these analyses only use a portion
of the full sample, and while the point estimates are largest for the middle
tercile, they are not statistically different from the effects for the lowest or
highest tercile.
Earnings has a similar effect size as the main results, but is significant
nowhere here, in contrast to Hurwitz and Smith (2016) who find that the
earnings effect of the Scorecard was strongest for students with the highest
SAT scores. I do not replicate that result here, and instead find no statistical
difference. The difference may arise because they compare high and low SAT
students and high schools, as opposed to high and low SAT colleges. If high-
SAT students are a relatively small proportion of the sample, their response
may not be visible in the aggregate college-level data.
Looking at colleges by the proportion of their students who receive Pell
grants addresses an important finding in the prior literature. Previous studies
typically find that students with low incomes are most responsive to new
information, perhaps because they have the least information to begin with.
However, the literature also finds that students with low incomes are the
least likely to seek out information, such as the College Scorecard, in the
first place.
While the tuition effect is concentrated among high-Pell recipient colleges,
the graduation rate and earnings effects are much stronger at colleges with
low and medium levels of Pell recipients. These results are rather surpris-
ing. The fact that tuition has its strongest effects among high-Pell colleges
suggests that those interested in high-Pell colleges do seem to be finding and
responding to information that those interested in lower-Pell colleges are not,
which is unexpected. Given that there is some response to information, it
is surprising again that the result is not replicated for the other attributes,
when prior literature finds that low-income students respond more strongly
to new information in general. Importantly, looking at the effects within
groups of colleges based on the types of students they typically attract is not
the same as looking at the effects within particular types of students, and
this distinction may explain some of these unusual results.
The division in the Scorecard’s effect by the primary degree awarded is
perhaps the most stark. The effects of the Scorecard are be driven almost en-
tirely by colleges that primarily award Bachelors’ degrees (although the effect
is small even for these colleges, and earnings has an unintuitive sign in this

22
subgroup). These results make sense if those interested in community college
are less likely to search for college information, or are more geographically
constrained and thus have less reason to compare colleges.
The results in this section suggest that the effect of the Scorecard on ag-
gregate search behavior is concentrated in BA-granting colleges, and colleges
with medium to large enrollments. The particular college attributes that the
Scorecard seems to boost differ based on the different types of colleges.
This analysis so far ignores selection between different types of colleges.
If searchers respond to the Scorecard information by switching from AA-
granting colleges to BA-granting colleges rather than to other AA-granting
colleges with better attributes, this will not show up in Table 4. So, in Table
5, I show how the Scorecard information affected interest in colleges by their
type, rather than by their attributes.
The first model shows how the Scorecard affects interest in colleges by
college control and degree awarded. This distinguishes between six types
of colleges: public, private non-profit, and for-profit, each in primarily-AA-
granting and BA-granting varieties. Public AA-granting colleges are the
reference group. The Scorecard does seem to have a significant impact
on interest in different college types. In particular, the Scorecard seems
to be directing interest away from AA-granting colleges and towards BA-
granting colleges. This offers one potential explanation for the lack of a
Scorecard effect among AA-granting colleges, if those who viewed the infor-
mation searched instead for BA-granting institutions rather than choosing
differently between AA-granting institutions, or if the Scorecard only incited
search activity among those searching for BA colleges and had no effect on
those looking at AA colleges. There is also evidence that the Scorecard
seems to direct interest away towards not-for-profit BA colleges, relative to
for-profit BA colleges. There is no significant effect for for-profit AA colleges
relative to other AA colleges.
While these effects, like all the other effects of the Scorecard found in
the paper, are meaningfully small, it is interesting that extending further
information drives student interest away from for-profit BA colleges. Since
the Scorecard does report for-profit status, it is plausible that this effect
represents a small number of students who prefer not-for-profit colleges and
learn from the Scorecard which colleges are for-profit.
The second model shows how interest in colleges by enrollment is affected,
with large colleges as the reference group. The third divides colleges by test
scores, with high-score colleges as the reference group. These models both

23
Table 5: Effects of the College Scorecard on College Type Interest

(1) (2) (3) (4)

Pub. BA-Granting 2.870***
× Scorecard (.231)
Priv. AA-Granting -.420
× Scorecard (.433)
Priv. BA-Granting 1.241***
× Scorecard (.200)
For-Profit AA -.309
× Scorecard (.253)
For-Profit BA -.673*
× Scorecard (.357)
Enrollment < 2,000 -3.678***
× Scorecard (.466)
Enrollment 2,000-15,000 -2.455***
× Scorecard (.470)
Lowest SAT Tercile -1.002**
× Scorecard (.442)
Medium SAT Tercile -.622
× Scorecard (.423)
Lowest Pell Tercile 4.115***
× Scorecard (.285)
Medium Pell Tercile 3.338***
× Scorecard (.297)
N 127,206 127,095 51,245 127,095
All models include college fixed effects and use an unweighted index.
*/**/*** indicates statistical significance at the 10%/5%/1% level.

24
show interest shifting towards high-enrollment and high-test-score colleges,
respectively.
The fourth model shows how interest in colleges with different levels of
Pell recipients was affected by the Scorecard. The Scorecard seems to have
increased interest in colleges with low levels of Pell recipients relative to
high levels. This could be interpreted as shifting attention towards low-Pell
colleges, but also may reflect the type of person using the Scorecard, and the
type of college they are likely to be interested in. People who would have
been interested in low-Pell colleges being more likely to use the Scorecard,
and searching more heavily in general as a result would also produce this
result.

5.3 Local Average Treatment Effects

Table 6 shows the results of the instrumental variables regression described
in Section 4. Each attribute is evaluated separately since there is only one
instrument, with the note that the result for tuition should be considered with
caution, given that it reverses sign when all three attributes are included in
the same model in Table 1.
The first stage of the regression is strong in each case. Unsurprisingly,
the release of the Scorecard itself significantly increases search intensity for
“college scorecard.” This jump in interest can be seen in Figure 2. The second
stages have signs that match the individual-attribute regressions in Table 2.
The magnitudes have changed, but keep in mind that these coefficients now
represent the sum total effect of all compliers.
As mentioned in Section 4, an estimate of the LATE can be derived by
dividing the coefficient on the second-stage interaction term (the total effect
of all compliers) by N , the number of people each unit of ScorecardInterestt
represents. I calculate N using website hits for the College Scorecard. From
April 13-May 12, 2016, the College Scorecard was visited by 56,974 unique
users. During this same period, the Google Trends index for “college score-
card” is 4, from a search window that covers the past three years.
The Scorecard index is scaled so that the maximum and minimum search
interest in a given period lie between 0 and 1. Making the simplifying as-
sumption that the minimum search activity was a true zero, a single unit on
the index translates into 56, 974/4 = 14, 243.5 users. This is an inexact es-
timate, since the Google index reports only integers and not decimal points,
and since scale may change somewhat over time as the total number of all

25
Table 6: Local Average Treatment Effects using IV

Variable College Scorecard

Interest
2SLS Second Stage for Tuition
Tuition ($1k) × .004***
ScorecardInterestt (.001)
N 121,582
First Stage
Scorecard 221.906***
(1.175)
2SLS Second Stage for Earnings
Earnings ($1k) × .002***
ScorecardInterestt (.000)
N 117,142
First Stage
Scorecard 500.031***
(2.468)
2SLS Second Stage for Grad. Rate
Grad. Rate × .184***
ScorecardInterestt (.021)
N 113,664
First Stage
Scorecard 5.806***
(.032)
All models include college fixed ef-
fects and a quadratic time trend.
ScorecardInterestt is a Google Trends
index for the term “college scorecard.”
*/**/*** indicates statistical signifi-
cance at the 10%/5%/1% level.

26
 Figure 2: Google Trends Results for “College Scorecard”

Google searches changes Stephens-Davidowitz and Varian (2014). However,

it is likely a reasonable approximation since most search activity for the
Scorecard occurred over a relatively short recent window after its release.
I divide each of the coefficients by 14,243.5, focusing on Earnings and
Graduation rate for these individual effect models. The coefficient on earn-
ings then becomes 1.4 × 10−7 , and the coefficient on graduation rate becomes
1.3×10−5 . These are the estimates of β1 in Equation 2. The treatment of one
additional complier would increase the Google Trends index for a college by
the product of β1 and the attribute. One additional person treated because
of the release of the Scorecard would change aggregate behavior in such a way
that a college with $40,000 earnings would see an increase in their Google
Trends index of 1.4 × 10−7 × (40 − 35) = 7 × 10−7 more than a college with
$35,000 earnings. Comparing a college with a graduation rate of 60% to one
with 50%, the effect would be 1.3 × 10−5 × (.6 − .5) = 1.3 × 10−6 . In each
case the effect is approximately one ten-millionth of a standard deviation of
search behavior.
Again, it should be kept in mind that these figures rely on the estimate of
N , the number of treated users, which is approximate. But under the current
estimate, it would take an additional 1,000,000 users such that meaningfully
different colleges would see an improvement of even .1 of a standard devi-
ation. If N is overestimated by several orders of magnitude, this could be

27
an impressive LATE. But the more N is overestimated, the farther away the
Scorecard currently is from its goal. Overall, whatever the current number of
users is, the user base would have to be increased by a factor of 176 in order
for the Scorecard to have an effect of .1 of a standard deviation comparing
colleges with $5,000 income differences or 10% graduation rate differences.

6 Conclusion
This paper presents estimates of the effect of the College Scorecard on search
behavior. The Scorecard led people to search more often for high-earnings,
high-graduation rate, and low-tuition colleges, as intended. However, whether
measured as the aggregate intent-to-treat effect of the Scorecard on search
behavior, or as the local average treatment effect of one additional person us-
ing the website, the effect was extremely small. The Scorecard has only small
effects on search patterns or aggregate search activity, and would need to re-
cruit many more users in order to make a meaningful impact on aggregate
search behavior and drive more interest towards high-performing colleges.
These results largely match those of Hurwitz and Smith (2016), who study
the effects of the Scorecard on student SAT-sending behavior.
It is worth emphasizing that affecting student behavior is not the only
goal of the College Scorecard website, which also has uses for federal funding
decisions and applications for researchers. Given the low marginal cost of
providing information to students after these other needs have been met, the
small effect sizes may be considered acceptable. Still, the policy is not having
a major impact on its intended audience.
Evidence on other information-based interventions means the small effect
of the Scorecard should have been predictable. Other studies of information-
only interventions in college education, in which students are informed about
the costs or benefits of different levels or types of education, tend to find small
or null effect sizes.
The biggest successes of informational interventions in education tend to
come when the policy targets low-income students or regions, where the infor-
mation deficit may be largest (see Avery and Kane 2004; Rouse 2004; Hast-
ings et al. 2016; Huntington-Klein 2016b on information deficiencies among
low-income students, and Nguyen 2008; Jensen 2010; Hoxby and Turner 2013;
Hastings et al. 2015 on informational interventions targeting low-income stu-
dents or regions). However, the Scorecard, rather than informing students

28
directly, makes information available on the internet, where it is likely that
better-off students are more likely to access it. This may explain why Hurwitz
and Smith (2016) find the largest effects of the Scorecard among students
from privileged backgrounds, and I find that earnings and graduation rate
have the strongest effects on search behavior for colleges with lower levels of
Pell recipients.
In general, even among intensive interventions that directly provide in-
formation to students, we still do not know yet how to structure information-
only interventions such that they all produce results as large as, say, Hoxby
and Turner (2013), who manage to change student college application be-
havior by amounts as large as a half of a standard deviation. The Scorecard
is a much less intensive intervention than even many of these only mildly
successful studies, and so is fighting an uphill battle. It did not buck the
disappointing trend.
However, the previous literature may provide some cause for optimism.
Given that the Scorecard does work as intended, just not to the same degree
as intended, increasing usage through marketing and, potentially, targeting
usage, would increase the Scorecard’s aggregate effect. With the current
estimate of the LATE, the site would have to attract at least a hundred
times as many users to have a meaningful impact. But, leaning on the
literature’s general finding that low-income students respond more strongly
to information, the effect of each user may increase if the Scorecard can
focus on recruiting low-income users. These students are likely to have more
need of the information in the first place, and respond more strongly to the
availability of the information if they know about the Scorecard and are
encouraged to use it.
Finally, aside from any result about the College Scorecard in particular,
this study outlines an approach to using Google Trends data in policy analysis
more broadly. In any case where public interest or opinion is a policy-relevant
variable, Google Trends may be used to capture that data in a way that does
not rely on infrequent surveys with relatively small sample sizes. There are
downsides to this approach as well; Google Trends indices can require work
to make them comparable across time or across keywords, and the estimated
effect size is not easily interpretable in absolute terms. But these are tradeoffs
with the downsides that surveys themselves necessarily have. Just as it has
previously been used in prediction, Google Trends data can also be used in
the estimation of causal effects.

29
References
Attanasio, Orazio P., and Katja M. Kaufmann. 2009. Educational Choices,
Subjective Expectations, and Credit Constraints.

———. 2012. Education Choices and Returns on the Labour and Marriage
Markets: Evidence from Data on Subjective Expectations.

Avery, Christopher, and Thomas J. Kane. 2004. Student Perceptions of

College Opportunities: The Boston COACH Program. In College choices,
ed. Caroline Hoxby, 355–394. Chicago, IL: University of Chicago Press.

Bergman, Peter, Jeffrey T. Denning, and Dayanand Manoli. 2016. Is Infor-

mation Enough ? Evidence from a Tax Credit Information Experiment
with 1,000,000 Students.

Bettinger, Eric, Bridget T. Long, Philip Oreopoulos, and Lisa Sanbonmatsu.

2012. The Role of Simplification and Information in College Decisions:
Results from the H&R Block FAFSA Experiment. The Quarterly Journal
of Economics 127(3):1205–1242.

Betts, Julian R. 1996. What do Students Know about Wages? Evidence from
a Survey of Undergraduates. Journal of Human Resources 31(1):27–56.

Booij, Adam S., Edwin Leuven, and Hessel Oosterbeek. 2012. The Role of
Information in the Take-up of Student Loans. Economics of Education
Review 31(1):33–44.

Botelho, Anabela, and Ligia C. Pinto. 2004. Students’ Expectations of the

Economic Returns to College Education: Results of a Controlled Experi-
ment. Economics of Education Review 23(6):645–653.

Buss, Christian, Jeffrey Parker, and Jon Rivenburg. 2004. Cost, Quality and
Enrollment Demand at Liberal Arts Colleges. Economics of Education
Review 23(1):57–65.

Castleman, Benjamin L., and Lindsay C. Page. 2015. Summer Nudging: Can
Personalized Text Messages and Peer Mentor Outreach Increase College
Going Among Low-income High School Graduates? Journal of Economic
Behavior and Organization 115:144–160.

30
Choi, Hyunyoung, and Hal Varian. 2012. Predicting the Present with Google
Trends. Economic Record 88(SUPPL.1):2–9.
Dominitz, Jeff, and Charles F. Manski. 1996. Eliciting Student Expectations
of the Returns to Schooling. The Journal of Human Resources 31(1):1–26.
Dynarski, Susan. 2003. Does Aid Matter? Measuring the Effect of Student
Aid on College Attendance and Completion. American Economic Review
93(1):279–288.
Fryer, Roland G. 2017. Information, Non-Financial Incentives, and Student
Achievement: Evidence from a Text Messaging Experiment. Journal of
Public Economics (Forthcoming).
Griffith, Amanda, and Kevin Rask. 2007. The Influence of the US News
and World Report Collegiate Rankings on the Matriculation Decision of
High-ability Students: 1995-2004. Economics of Education Review 26(2):
244–255.
Grodsky, Eric, and Melanie T Jones. 2007. Real and imagined barriers to
college entry: Perceptions of cost. Social Science Research 36(2):745–766.
Hastings, Justine S., Christopher A. Neilson, Anely Ramirez, and Seth D.
Zimmerman. 2016. (Un)informed College and Major Choice: Evidence
from Linked Survey and Administrative Data. Economics of Education
Review (Forthcoming).
Hastings, Justine S., Christopher A. Neilson, and Seth D. Zimmerman. 2015.
The Effects of Earnings Disclosure on College Enrollment Decisions.
Hoxby, Caroline, and Sarah Turner. 2013. Expanding College Opportunities
for High-Achieving, Low Income Students.
Huntington-Klein, Nick. 2015. Subjective and Projected Returns to Educa-
tion. Journal of Economic Behavior & Organization 117:10–25.
———. 2016a. A Method for Estimating Local Average Treatment Effects
in Aggregate Data with Imperfect Assignment. Applied Economics Letters
(Forthcoming).
———. 2016b. ”(Un)informed College and Major Choice”: Verification in
an Alternate Setting. Economics of Education Review 53:159–163.

31
Hurwitz, Michael, and Jonathan Smith. 2016. Student Responsiveness to
Earnings Data in the College Scorecard.

Imbens, Guido W., and Joshua D. Angrist. 1994. Identification and Estima-
tion of Local Average Treatment Effects. Econometrica 62(2):467–475.

Jacob, Brian A., Brian McCall, and Kevin M. Stange. 2016. College as Coun-
try Club: Do Colleges Cater to Students’ Preferences for Consumption?
Journal of Labor Economics (Forthcoming).

Jensen, Robert. 2010. The (Perceived) Returns to Education and the Demand
for Schooling. The Quarterly Journal of Economics 125(2):515–548.

Kerr, Sari Pekkala, Tuomas Pekkarinen, Matti Sarvimaki, and Roope Uusi-
talo. 2015. Post-Secondary Education and Information on Labor Market
Prospects: A Randomized Field Experiment.

Kodde, David A. 1987. Unemployment Expectations and Human Capital

Formation. European Economic Review 32(8):1645–1660.

Manski, Charles F. 1993. Adolescent Econometricians: How Do Youth Infer

the Returns to Schooling? In Studies of supply and demand in higher
education, ed. Charles T. Clotfelter and Michael Rothschild, chap. 2, 43–
57. Chicago, IL: University of Chicago Press.

Monks, James, and Ronald G. Ehrenberg. 1999. The impact of US News

& World Report College Rankings on Admissions Outcomes and Pricing
Policies at Selective Private Institutions. Change: The Magazine of Higher
Learning 31(6):42–51.

Nguyen, Trang. 2008. Information, Role Models and Perceived Returns to

Education: Experimental Evidence from Madagascar.

Oreopoulos, Philip, and Ryan Dunn. 2013. Information and College Access:
Evidence from a Randomized Field Experiment. The Scandinavian Journal
of Economics 115(1):3–26.

Oreopoulos, Philip, and Uros Petronijevic. 2013. Making College Worth It:
A Review of Research on the Returns to Higher Education. The Future of
Children 23(1):41–65.

32
Oreopoulos, Philip, and Kjell G. Salvanes. 2011. Priceless: The Nonpecu-
niary Benefits of Schooling. The Journal of Economic Perspectives 25(1):
159–184.

Parker, Jeffrey, and Jeffrey Summers. 1993. Tuition and Enrollment Yield
at Selective Liberal Arts Colleges. Economics of Education Review 12(4):
311–324.

Rouse, Cecilia Elena. 2004. Low-Income Students and College Attendance:

An Exploration of Income Expectations. Social Science Quarterly 85(5):
1299–1317.

Smith, Herbert L., and Brian Powell. 1990. Great Expectations: Variations
in Income Expectations among College Seniors. Sociology of Education
63(3):194–207.

Stephens-Davidowitz, Seth. 2014. The Cost of Racial Animus on a Black

Candidate: Evidence using Google Search Data. Journal of Public Eco-
nomics 118:26–40.

Stephens-Davidowitz, Seth, and Hal Varian. 2014. A Hands-on Guide to

Google Data. Tech. Rep., Google.

Stinebrickner, Ralph, and Todd R. Stinebrickner. 2014. A Major in Science?

Initial Beliefs and Final Outcomes for College Major and Dropout. Review
of Economic Studies 81(1):426–472.

Usher, Alex. 2005. A Little Knowledge is a Dangerous Thing: How Per-

ceptions of Costs and Benefits Affect Access to Education. Tech. Rep.,
Toronto.

Wiswall, Matthew, and Basit Zafar. 2015a. Determinants of College Ma-

jor Choice: Identification using an Information Experiment. Review of
Economic Studies 82(2):791–824.

———. 2015b. How Do College Students Respond to Public Information

about Earnings? Journal of Human Capital 9(2):117–169.

Zafar, Basit. 2011. How Do College Students Form Expectations? Journal

of Labor Economics 29(2):301–348.

33